Fault diagnosis of power transformer using optimally selected DGA features and SVM

Abstract

Dissolved Gas Analysis (DGA) is an established method for detecting and predicting faults contained in power transformers. Support Vector Machine (SVM) has been actively applied to classify faults using historic DGA data. However, redundant and irrelevant features can reduce SVM classification performance. Therefore, this study proposes the use of GA-SVM wrapper to eliminate these features and select optimal features from DGA dataset to be used as inputs to SVM. GA-SVM wraps Genetic Algorithm (GA) around SVM, meaning that the estimated accuracy of SVM becomes the fitness function for each of the subsets found or generated by GA. Using these optimal features, SVM is trained and tested using two different datasets. The accuracies of SVM learned on the full set of features and that learned on the selected subsets by GA are compared using two real-world DGA datasets. Experimental results show that SVM performs better using optimal DGA subset than the whole dataset. It can be concluded that the proposed method which combines GA-SVM and SVM eliminates redundant features and improves SVM performance in classifying transformer fault based on DGA data.